TL;DR
This paper provides a comprehensive survey and empirical evaluation of adversarial attacks specifically targeting object detection models. It introduces a novel taxonomy for these attacks, analyzes existing evaluation methodologies, and offers crucial insights into the vulnerabilities of various object detectors, highlighting the need for standardized evaluation and more robust defense mechanisms.
What Problem Does This Paper Address?
Deep learning models have revolutionized computer vision, but they are susceptible to adversarial examples – subtle input perturbations that cause incorrect predictions. While adversarial attacks on image classification have been extensively studied, research on object detection remains limited. This is a critical gap because object detection is vital for high-stakes applications like autonomous vehicles and security surveillance. Existing surveys and evaluations suffer from limitations such as narrow scopes, lack of open-source code, and incompatible dependencies, making it difficult to compare attack effectiveness across different object detection architectures. This paper aims to bridge this gap by providing a unified framework and rigorous empirical analysis.
Methods
The authors propose a novel taxonomy to categorize adversarial attacks in object detection, classifying them based on:
- Adversary Model: Digital vs. Physical attacks.
- Knowledge: White-box, Gray-box, and Black-box attacks, detailing the adversary's access to model information.
- Intent Outcome: Integrity-based attacks (miscategorization, object vanishing, object fabrication) and availability-based attacks (inference latency).
- Detector Type: Anchor-based (one-stage, two-stage) and Anchor-free (one-stage, two-stage) detectors.
- Component Under Attack: Classification confidence, objectness score, shape regression, relevance map, features, context, region proposal, and Non-Maximum Suppression (NMS).
- Perturbation Norm: Lp-norm, PSNR, SSIM.
- Frequency: One-shot vs. Iterative attacks.
- Attack Specificity: Image-specific vs. Universal attacks.
- Region Specificity: On-object, On-background, Unrestricted.
- Locality: Patch vs. Entire image attacks.
The paper also synthesizes existing evaluation metrics and proposes a comprehensive empirical evaluation of state-of-the-art open-source attack methods on popular object detection models.
Datasets
The paper uses the following datasets for evaluation:
- COCO 2017 validation set: Used for evaluating attack effectiveness and transferability studies. A subset of 500 images was used for ensemble attacks to reduce evaluation time.
- PASCAL VOC: Mentioned as a common dataset in the literature.
- KITTI, Inria Person, CityPersons, ImageNet: Mentioned as datasets used in other literature.
- Mapillary, Traffic Sign: Mentioned as datasets used in other literature.
Key Insights and Ideas
- No Correlation Between White-box and Black-box Robustness: Surprisingly, detectors that are robust against white-box attacks are not necessarily robust against black-box attacks, and vice-versa. This highlights the distinct challenges posed by different attack knowledge levels.
- Smaller Objects are More Vulnerable: Adversarial attacks disproportionately affect the detection of small objects, showing significantly higher mAP drops compared to medium or large objects. This suggests a specific area of concern for future defense research.
- Modern Architectures Show Improved Resistance: Newer, more complex detector architectures (like Transformer-based models such as DeTR and GLIP) generally exhibit better robustness against adversarial attacks compared to traditional detectors.
- Diverse Surrogate Ensembles Boost Transferability: Using a diverse group of surrogate detectors for black-box attacks significantly increases their transferability and effectiveness against unseen victim models.
- Patch Attacks Can Be Visually Obvious: While some patch attacks can achieve significant mAP drops, the perturbed patches are often easily recognizable by human eyes, limiting their practical applicability in real-world scenarios.
Key Results
- TOG is identified as an effective white-box attack with a low time cost, achieving mAP drops of over 50% with minimal queries.
- EBAD demonstrated impressive performance in ensemble-based black-box attacks, achieving mAP drops of over 22.55% within 10 queries, and its single-query performance also outperformed other methods.
- DPatch was found to be ineffective in the paper's experimental setup with limited queries, likely due to its original design requiring a large number of iterations.
- Cross-model black-box attacks showed that detectors like DeTR and GLIP had the lowest mAP drops (around 8-15%), indicating higher robustness.
- Transferability studies revealed that smaller objects are more vulnerable, with mAP drops for small objects being 10-20% higher than for larger objects across various detectors.
Limitations and Future Work
- Limited Availability of Open-Source Code: Some attack methods were excluded due to outdated or unmaintained code, or lack of clear setup instructions, hindering comprehensive evaluation.
- Standardized Evaluation Needed: The paper emphasizes the urgent need for standardized evaluation protocols to enable fair comparisons across different attack methods and detector architectures.
- Research on Small Object Defenses: The vulnerability of small objects to adversarial attacks is a significant gap, requiring dedicated research into both attack and defense strategies.
- Image Rendering Attacks Under-Explored: Adversarial attacks based on image rendering (rotation, translation, lighting) are less explored for object detection and hold practical implications for applications like autonomous driving.
- Multimodal Model Robustness: The increasing use of multimodal models (e.g., RGB, LiDAR, depth) introduces new attack surfaces that require further investigation for robust defense mechanisms.
- Physical Attack Evaluation: Evaluating physical adversarial attacks is challenging due to cost and logistical complexities, necessitating better methods for ensuring evaluation fidelity and reliability.
Summary updated